The identification of factors explaining bank profitability has been the subject of modeling in the literature. However, given the differences in the contexts of these studies, not all of these models are appropriate for explaining bank profitability in the WAEMU. The specification of our model is based on the work of [21] and [7]. Building on the study by [21], [7] specified the following model of bank profitability:

where, MI n t is the banking intermediation margin; α n is the constant term; j = {IPF, TPIBR, INFL, and RESTA}.

In this specification, CCTA refers to the ratio of customer loans to total assets on banks’ balance sheets; DCTA refers to the ratio of customer deposits to total assets on banks’ balance sheets; FGTA refers to the ratio of operating expenses to total assets on banks’ balance sheets; CDTA refers to the ratio of non-performing loans to total assets on banks’ balance sheets; TEMR denotes the real average discount rate; IPF denotes the financial policy index; TPIBR denotes the real per capita gross domestic product growth rate; INFL denotes the inflation rate; and RESTA denotes the ratio of reserves to total balance sheet assets ([7]).

This specification of the bank profitability model pertained to the WAEMU. We therefore adjust the previous model to the context of our study by taking into account credit risk and political risk. We arrive at the following bank profitability model:

where, MGO j t is the aggregate margin of banks in country j at time t; CRISK j t is a measure of credit risk; PRISK j t is a weighted index of political risk; α j is the constant term; and u j t are the residuals.

The WAEMU countries included in our study are: Burkina Faso, Côte d'Ivoire, Guinea-Bissau, Mali, Niger, Senegal, and Togo. Only Benin is not included in the sample due to a lack of data, particularly regarding political risk. However, the exclusion of Benin does not limit the study’s results. The study uses national-level panel data. We defined the study period (2005-2023) based on data availability. Since Benin does not present a particularly heterogeneous case with regard to the subject of our study, these seven countries are representative of all WAEMU countries (Table 1).

Table 1. Description of the variables in Model 1.

Source: Author.

In this section, we conducted alternative unit root and cointegration tests to examine the stationarity of the various variables in the model and to identify the appropriate estimator for this study.

In this study, we conducted the LLC unit root tests by [44]; the IPS tests by [34]; and the test by [30]. These tests are robust and frequently used in the literature. The panel unit root test by [44] assumes a homogeneous autoregressive unit root, whereas that by [34] assumes a heterogeneous unit root.

We have summarized the results of these various unit root tests at the level and first-difference levels in Appendix 2 and Appendix 3, respectively. These results from the panel unit root tests are consistent and show that all variables are integrated of order 1. Indeed, based on the IPS unit root tests, all variables in the model are first-order integrated. However, based on the results of the IPS tests and LLC unit root tests, all variables are integrated of order 1 except the RESTA variable. According to [30] tests, the TPIBR, INFL, and RESTA variables are not stationary in first differences.

Given the small sample size (N = 7; T = 19) and the potential heterogeneity among WAEMU countries, the IPS tests appear to be the most appropriate among first-generation tests. This test allows for heterogeneous dynamics among panel units and exhibits better properties in small samples compared to the LLC test. The Hadri test is used as a complementary robustness check. We therefore gave priority to the results of this test.

Based on the results of the previous unit root tests, we conducted alternative panel cointegration tests. In principle, to test for the existence of cointegration vectors among the variables under consideration, the traditional approach of [59] appears to be the best option. However, given the small size of our sample, we opted for [38] continuity test. Indeed, using Monte Carlo simulations, [29] showed that the smaller the sample size, the more Kao’s tests outperform those of Pedroni. In the specific case of our research, Kao’s cointegration tests are therefore more appropriate. We have summarized the results of these cointegration tests in the appendix (see Appendix 4). The various results of Kao’s cointegration tests reject the null hypothesis of non-cointegration of the variables in the different models. We therefore consider the different series to be cointegrated.

The previously specified bank profitability model uses macroeconomic variables. Consequently, certain variables in the model may be jointly determined and lead to potential endogeneity issues. As a result, estimating the model using the Ordinary Least Squares method may yield biased results. Given these concerns and the results of the cointegration tests, estimating our model using the Pooled Mean Group (PMG) method by [63] or the Fully Modified Ordinary Least Squares method by [60] is more appropriate. We chose the PMG model because it allows us to reconcile the assumption of a common long-run relationship among the panel units with the existence of differentiated short-run adjustments. Unlike the MG model, it improves the efficiency of the estimates by requiring homogeneity of the long-run coefficients, while remaining more flexible than the DFE model, which assumes complete homogeneity of the coefficients.

The PMG method allows for reconciling, within a single specification, the routine approach that imposes fixed coefficients and the approach that assumes country-specific coefficients. This implies that the long-run relationship between the variables is identical for all countries, but that each country follows its own dynamics to converge toward this common relationship. This approach appears to be the most appropriate in the case of the WAEMU countries.

According to [63], model (1) can be viewed as an autoregressive model with lagged variables (ARDL) of the form:

where X i t is the vector of explanatory variables, and µ i represents the fixed effect (country). If the variables are cointegrated, then the error term ε i t is a stationary process. In this case, the model can be re-specified as an error correction model in which the short-run dynamics are influenced by the deviation from the long-run relationship:

where θ i is the vector of long-term coefficients and ∆ is the operator of change between two successive dates. The adjustment coefficient, ϕ i , and the long-run coefficients, θ i , constitute the parameters of interest.

One of the advantages of ARDL models is that the short-run and long-run multipliers are estimated jointly. Furthermore, these models allow for the presence of variables that may be integrated of different orders, namely I(0) and I(1), or cointegrated ([63]).

The coefficient of the error correction term (COINTEQ01) is negative and statistically significant (−0.744716; p-value = 0.0000), confirming the existence of a long-run relationship between the variables. This coefficient indicates that approximately 74.47% of the imbalance observed in the previous period is corrected in each period, reflecting a rapid rate of adjustment toward long-run equilibrium. In this study, we conducted three alternative unit root tests on the data. These are the tests by [44], [34], and [30]. These alternative tests allowed us to verify the robustness of the stationarity analysis results for the series. We also conducted two alternative cointegration tests on the study data. These are the cointegration tests by [38] and [60]. These cointegration tests allowed us to verify the existence of a long-run relationship between the variables in our bank profitability model. Following these various unit root and cointegration tests, we alternatively used two estimation techniques: the PMG estimator ([63]) and the FMOLS estimator ([60]). These two estimation approaches allowed us to verify the stability of the estimated coefficients and the robustness of the estimation results. We used the FMOLS estimator as a robustness test for the PMG. Although FMOLS does not systematically address potential cross-country autocorrelation, it is widely used in macroeconomic, financial, banking, and economic growth studies. Indeed, in a context similar to that of the WAEMU, it corrects for endogeneity issues, autocorrelation of the residuals, and improves the accuracy of long-run estimates. It is highly effective for small samples ([60]) and when the variables are first-order integrated.

The estimated coefficients of the “CRISK” variable are significant and negative in all estimates. Credit risk, therefore, has a negative long-term effect on bank profitability in the WAEMU. Indeed, in the WAEMU, credit risk not only leads to difficulties in loan recovery but also causes banks to distance themselves (mistrust) from borrowers. It consequently increases banks’ cost of capital, thereby reducing bank margins. [14] found a similar result regarding bank profitability in the case of certain Latin American countries. This result also corroborates the findings of [32] in the context of Taiwan region.

The estimated coefficients for the “PRISK” variable are significant and negative in all estimates. In accordance with the methodology, the interpretation of the estimated PRISK coefficient differs from the coding of PRISK values (scores). Indeed, a high score PRISK reflects a low political risk, and a low score reflects a high political risk.

However, in econometric models, a positive coefficient for this index indicates a positive effect, and a negative coefficient indicates a negative effect. Political risk, therefore, has a negative long-term effect on bank profitability in the WAEMU. Indeed, WAEMU countries are exposed to high political risks that constrain banking activity and undermine bank profitability in this region. Studies on bank profitability have found similar results. Indeed, in Turkey, political risk has a negative effect on bank profitability ([66]). [76] also found that a change in government harms bank profitability in Pakistan. [75], on the other hand, found that political instability has a positive and significant impact on the profitability of Yemeni banks (Table 2).

Table 2. Results of model estimates using the PMG method.

*Note: p-values and any subsequent tests do not account for the model selection.

When estimating our bank profitability model using the FMOLS method, we obtain results similar to the previous ones, particularly regarding interest rate variables. The estimated coefficients for credit risk (CRISK) and political risk (PRISK) are significant and negative. These variables have negative long-term effects on bank profitability in the WAEMU. Table 3 below provides the results of this method. However, the results of the PMG approach are more robust compared to those of the FMOLS approach.

Table 3. Results of model estimates using the FMOLS method.

According to various estimates using the PMG technique, loans, deposits, reserves, and bank overhead costs significantly affect bank profitability in the WAEMU.

The estimated coefficients for the “CCTA” variable (customer loans) are significant and positive. Bank credit, therefore, has a positive effect on bank profitability in the WAEMU. This result is consistent with economic theory. However, [21] found a negative effect of bank credit in their analysis of bank profitability across 80 countries. The same is true for [7]. According to him, this seemingly paradoxical result can be explained in this monetary zone. He argues that banks have not passed on the decline in the discount rate to lending rates. The estimated coefficients of the customer deposits variable “DCTA” are significant and negative. This result can be explained by the high costs of managing bank deposits due to the segmentation of the financial system in the WAEMU. An increase in these costs reduces the bank’s margin. The findings of [21] and [7] corroborate this result. The coefficients of the variable measuring bank reserves “RESTA” are significant and positive. This result is contrary to our expectations. Banks pass on the costs associated with reserve requirements to their margins. The coefficients of the variable measuring overhead costs “FGTA” are significant and positive. This result is also contrary to our expectations. However, it is consistent with other empirical studies on this subject ([7]; [21]). Banks incorporate their overhead costs into the pricing of financial services. Indeed, the results for RESTA and FGTA can be explained by the fact that, in WAEMU countries, banks have pricing power. They therefore pass on the various costs of credit in their pricing. Similarly, the segmentation of the credit market allows banks to tailor their lending terms (rates, collateral, durations) to risk profiles.

The coefficients for the “TPIBR” economic growth rate are significant and negative. This result is similar to that of [7]. This finding may reflect one of the consequences of financial system segmentation in the WAEMU. Indeed, only a minority of the population has access to banking services in the WAEMU. The growth rate would therefore tend not to be a major factor in determining the bank margin.

The inflation rate “INFL” appears with significant and positive coefficients. Beyond inflation’s impact on banks’ interest rate risk, rising prices can be viewed as a factor of macroeconomic instability—that is, a key determinant of country risk. Consequently, a positive and significant coefficient suggests that inflation-related costs were less significant than the revenues generated by banks in the WAEMU. [14] also found a positive and significant coefficient for inflation on bank profitability. [64], however, found a significant and negative coefficient in determining the banking margin in the United States.

The figures correspond to the probabilities p. For p > 0.1, the null hypothesis of non-stationarity cannot be rejected according to the LLC and IPS tests. However, for p < 0.1, the null hypothesis of stationarity is rejected according to the Hadri test.

The figures correspond to the probabilities p. For p > 0.1, the null hypothesis of non-stationarity cannot be rejected according to the LLC and IPS tests. However, for p < 0.1, the null hypothesis of stationarity is rejected according to the Hadri test.

The values in parentheses are the p-values. (*), (**), and (***) indicate rejection of the null hypothesis of non-cointegration at the 10%, 5%, and 1% significance levels, respectively.